Rare-event experiments such as searches for neutrinoless double-beta decay run arrays of dozens to hundreds of germanium crystals. Each crystal has its own geometry, impurity profile and electronic noise. To reject background, these experiments classify each recorded pulse by its shape and keep only pulses that look like signal. The cut on that classifier score must be set separately for each crystal, because a single shared cut over-filters some detectors and under-filters others. Today each crystal's cut is fixed from its own calibration events. Calibration runs are infrequent, and the clean signal-like events they produce are few per detector, so these per-detector cuts are statistically noisy and go stale as detectors drift between calibrations.
Recommender systems faced the same structural problem and solved it differently. They do not estimate each user's threshold in isolation. They shift a shared model's cutoff using cheap side information about the user and regularize each personal threshold toward a global one. We propose to apply that idea to germanium pulse-shape discrimination. A per-channel cut on a shared classifier is learned with partial pooling across channels and conditioned on continuously logged channel telemetry such as baseline noise and leakage current. It is then calibrated to hold a target signal acceptance. Our claim is narrow and testable. When a channel has few calibration events, or the cut must be updated between calibrations, the pooled, telemetry-conditioned cut should keep each channel's realized signal acceptance closer to target and reject at least as much background as independent per-channel calibration. When calibration statistics are plentiful, the advantage should disappear.
What the two directions contribute, and what is already done
The personalized-decision-boundary literature supplies the mechanism. Three papers set thresholds for individuals on top of a shared scorer. Personalized Spam Filtering for Gray Mail moves a global spam model's threshold according to each user's historical reporting rate instead of training per-user models. This is the closest template for "shift the cut from cheap side information." Dynamic-K Recommendation with Personalized Decision Boundary learns a threshold $t_u$ for each user, regularized toward a global $t$. It names cold start for sparse users as the open problem, and that is exactly our sparse-calibration regime. When to Show a Suggestion? feeds a programmer identifier into a shared acceptance model and tunes suppression to a stated recall guarantee (94.7%). We borrow its idea of fixing the operating point by a guaranteed acceptance rather than an accuracy optimum.
Two richer formulations learn more boundary parameters but give no acceptance guarantee. Decision-Aware Preference Modeling for Multi-Behavior Recommendation learns boundaries per user, item and behavior. Advancing Loss Functions in Recommender Systems learns per-user truncation inside a robust loss. We deliberately do not adopt their parameter-per-entity design. With only tens of channels and few calibration events each, free per-channel parameters are the problem we are trying to fix.
The physics side supplies the application and its constraints. Deep learning based pulse shape discrimination for germanium detectors trains an autoencoder plus a 141-parameter head on thorium-228 calibration data. It uses double-escape-peak (DEP) events as a single-site, signal-like proxy and full-energy-peak (FEP) events as a multi-site, background-like proxy. It reaches 90% FEP rejection at 90% DEP acceptance, matching the classical $A/E$ discriminator, which is the ratio of peak current to energy. Machine learning-powered data cleaning for LEGEND uses one global classifier to separate physics pulses from electronics artifacts and reports per-channel manual tuning as slow and brittle. α-event characterization and rejection in point-contact HPGe detectors adds a delayed-charge-recovery cut for surface alphas and documents the drift that erodes static cuts. Together these papers motivate the gap: the classifiers are shared, but the cut is either global or set per detector from that detector's data alone.
The research file reports no work that connects the two directions. The obvious version of the connection, "give each channel its own threshold," is not new, though. Normalizing the cut per detector to a fixed DEP acceptance is standard physics practice. The new element is pooling across channels and conditioning on telemetry. That pooling is what addresses sparse calibration and drift between calibrations.
Core Claim
Let $s(x)$ be a shared pulse-shape score for waveform $x$. Each channel $c$ receives a cut $\tau_c$, chosen so that the fraction of that channel's DEP events above the cut equals a target acceptance $1-\epsilon$, here 90%.
The independent estimator takes the empirical quantile of channel $c$'s own $n_c$ DEP scores. Its acceptance error scales as $\sqrt{\epsilon(1-\epsilon)/n_c}$, which follows from the binomial variance of a quantile.
We instead model the DEP score distribution of each channel as a location-scale shift of a pooled distribution. The shift parameters $(\mu_c,\sigma_c)$ come from a hierarchical prior whose mean is a regression on channel telemetry $z_c$:
$$ \mu_c \sim \mathcal{N}(f_\mu(z_c), \omega_\mu^2), \qquad \tau_c = \hat\mu_c + \hat\sigma_c\, q_{\epsilon}, $$
Here $q_\epsilon$ is the $\epsilon$-quantile of the pooled standardized score, and $\hat\mu_c,\hat\sigma_c$ are the posterior means given channel $c$'s DEP events. For a channel with few events, the cut shrinks toward the telemetry-predicted value. For a well-sampled channel, it converges to the independent estimate.
Two further steps make the method usable in an experiment. First, a split-conformal correction on held-out DEP events from the same channels restores a finite-sample acceptance guarantee in the spirit of the recall guarantee of When to Show a Suggestion?. Second, between calibrations, $f_\mu$ and $f_\sigma$ are re-evaluated on current telemetry. This makes the cut track drift without a source run, analogous to the reporting-rate shift in Personalized Spam Filtering for Gray Mail.
Research Program
The shared scorer is held fixed, so that the comparison isolates how the cut is set. We use two scorers: the classical $A/E$, and an autoencoder-plus-head model reimplemented after Deep learning based pulse shape discrimination for germanium detectors.
We build three things. The first is the hierarchical cut-setting estimator with its conformal correction. The second is a telemetry regression that predicts each channel's score location and scale from logged quantities: baseline RMS, energy resolution at a reference line, leakage current, temperature, and time since the last calibration. The third is a drift protocol that predicts cuts for runs after a calibration and scores them against the next calibration.
The analysis question is when pooling helps. We vary the number of DEP events per channel, and we vary how far in time the prediction extends from the last calibration. The expected result is a crossover: pooling wins below some $n_c$, and independent calibration matches it above. The value of the work is in locating that crossover and showing it lies in the regime real arrays operate in.
Data are the main risk. The research file lists single-detector benches: the LEGEND Full Chain Test stand, the TUBE scanner, and thorium DEP/FEP calibration data. It contains no public multi-detector calibration set. The primary experiment therefore requires multi-detector thorium calibration data from a collaboration such as MAJORANA DEMONSTRATOR or LEGEND-200. If such access fails, the fallback builds an emulated array from single-detector calibration data. Channels are created by injecting measured noise levels and resolution degradations, and per-channel event counts are subsampled. A positive fallback result supports the statistical mechanism but not the telemetry-to-cut relationship, and we will label it that way.
Baselines
The naive reference is a single global cut on the shared score, chosen for 90% pooled DEP acceptance. It shows how much per-channel treatment matters at all. The research file's evidence that static global cuts mis-filter heterogeneous channels predicts it will have a wide spread of per-channel acceptance.
The strongest competitor is independent per-channel calibration. It sets each channel's cut at the empirical 90% DEP quantile from that channel's own calibration events, which is what experiments do now. It is strongest because it is unbiased and makes no modeling assumptions. With enough events it cannot be beaten on acceptance accuracy. It receives exactly the same calibration events, the same scorer and the same held-out evaluation events as our method.
The reported reference point is the aggregate performance of Deep learning based pulse shape discrimination for germanium detectors: 90% FEP rejection at 90% DEP acceptance, matching $A/E$. No reported result was found on Lacuna for per-channel acceptance spread under sparse calibration, or for cut accuracy between calibrations. We will measure both baselines ourselves.
Beating the strongest baseline means two things. At small $n_c$, the realized per-channel DEP acceptance on held-out events must have a lower root-mean-square deviation from 90%. At the same time, mean FEP rejection must stay within statistical error or improve, and the conformal guarantee must hold. Between calibrations, the telemetry-updated cut must track the next calibration's true cut better than the stale cut from the previous calibration.
Two ablations isolate the ingredients. Pooling without telemetry shrinks every channel toward the array mean. Telemetry without pooling uses $f_\mu(z_c)$ alone, without channel data. If pooling without telemetry captures nearly all the gain, the telemetry model contributes nothing for static cuts and matters only for drift tracking. We will report that outcome as such.
Why not just…?
Why not just keep calibrating each channel independently? This is the real threat. In the high-statistics regime we expect it to tie our method, because the posterior then converges to the empirical quantile. Our claim is limited to two regimes, few DEP events per channel and drift between calibrations. The acceptance-error-versus-$n_c$ curve will show whether those regimes overlap realistic operating conditions. If the crossover lies below the counts experiments routinely collect, the static-cut part of the proposal fails and only drift tracking remains.
Why not just feed the channel ID into the classifier, as with the programmer ID in When to Show a Suggestion?? A learned channel embedding moves the score but does not fix the operating point. A cut still has to be calibrated to 90% acceptance on each channel's sparse data, so the original problem returns. A new or reconfigured channel also has no embedding, which is the cold-start limitation flagged by Dynamic-K Recommendation. We include this variant as an additional baseline, with its cut set per channel, to test whether conditioning the score adds anything on top of pooled cut-setting.
Why not just run calibrations more often? Source runs cost live time and handling effort. More frequent calibration reduces both the sparsity problem and the staleness problem, and it is the simplest fix. We cannot price that cost from the file. What we can measure is how much prediction error builds up as a function of time since calibration. That curve tells an experiment how far it could stretch its calibration interval under our method compared with stale cuts.
Why not just use a contamination-robust training loss, like the Rényi loss of Advancing Loss Functions in Recommender Systems? FEP training samples do contain single-site events, so a robust loss could improve the shared score. That improvement is independent of how the cut is set. It changes $s(x)$, while our method changes $\tau_c$. We keep the scorer fixed so that the two effects do not mix, and we leave robust training to separate work.
Decisive Checks
The first and decisive check is the curve of per-channel acceptance error against DEP events per channel, comparing the pooled estimator with independent calibration. The pooled estimator should be clearly below at small $n_c$ and meet the independent curve at large $n_c$. Its empirical coverage should stay at or above 90% after the conformal correction.
The second check is the drift test. Cuts predicted from telemetry for runs after a calibration are compared, at the next calibration, against both the stale cut and that calibration's actual cut.
The third check guards against sculpting, the concern raised by Isolating Unisolated Upsilons that a cut can create artificial structure. We will verify that the per-channel cuts do not distort the energy spectrum near the signal region, using calibration lines away from the DEP as a control.
If pooling fails to beat independent calibration at the event counts real arrays collect, we will report that the standard practice is already adequate. That would be a useful negative result.
References
- Personalized Spam Filtering for Gray Mail
- Dynamic-K Recommendation with Personalized Decision Boundary
- When to Show a Suggestion? Integrating Human Feedback in AI-Assisted Programming
- Decision-Aware Preference Modeling for Multi-Behavior Recommendation
- Advancing Loss Functions in Recommender Systems: A Comparative Study with a Rényi Divergence-Based Solution
- Deep learning based pulse shape discrimination for germanium detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data